Open-source AI models you can run locally.
Updated Sep 11, 2026
Weights are not the product.
An open-weight model still needs a runner. Ollama, llama.cpp, llamafile, LocalAI, and vLLM are different ways to load the same class of GGUF or GPU weights.
Pick the runner for your OS and ops comfort first. Then pick a model that fits the RAM floor.
Size before brand.
1B–3B class models fit 8 GB RAM without a GPU if the desktop is quiet. 7B class wants more RAM or a GPU. Serving many users wants vLLM and real VRAM, not a laptop runner.
Fossfits lists runners and UIs. Model cards on Hugging Face still decide license and use limits for the weights themselves.